Int. J. Production Economics 144 (2013) 68–75 Contents lists available at SciVerse ScienceDirect Int. J. Production Economics journal homepage: www.elsevier.com/locate/ijpe The heterogeneous ﬂeet vehicle routing problem with overloads and time windows Manolis N. Kritikos n, George Ioannou Management Science Laboratory, Department of Management Science and Technology, Athens University of Economics and Business, Evelpidon 47A & Lefkados 33, 113-62 Athens, Greece a r t i c l e i n f o a b s t r a c t Article history: Received 30 September 2011 Accepted 18 January 2013 Available online 29 January 2013 This paper addresses a new variant of the vehicle routing problem (VRP) with time windows, in which the vehicle ﬂeet comprises units of different capacities and some overloads (i.e., loading vehicles above nominal capacity) are allowed. Although often encountered in practice, the problem, which we call heterogeneous ﬂeet vehicle routing with overloads and time windows (HFVROTW), has not been previously tackled in the literature. We model it by integrating the constraint of the total trip load into the objective function, and solve it via a sequential insertion heuristic that employs a penalty function allowing capacity violations but limiting them to a variable predeﬁned upper bound. Computational results on benchmark problems show the effectiveness of the proposed approach in reducing vehicle costs with minimal capacity violations, thus offering evidence of the signiﬁcance of this VRP variant. & 2013 Elsevier B.V. All rights reserved. Keywords: Vehicle routing Heuristics Time windows 1. Introduction Real-life vehicle routing emanates in a large variety of cases where goods or people have to be moved between locations in speciﬁc time intervals by different transportation means (Golden et al., 2008). The academic community has studied several versions of vehicle routing and scheduling problems, proposing an immense list of solution approaches, ranging from simple heuristics to complex meta-heuristics and exact methods—see Gendreau and Tarantilis (2010) for a thorough review of past and recent developments in the ﬁeld. Most researchers though, have concentrated on simpliﬁed instances of the problem, i.e., instances that do not accommodate constraints or objectives often encountered in practice, which guide the actual solutions in real vehicle routing and scheduling problems—see Kritikos and Ioannou (2010) for such a discussion. In this work, we study the heterogeneous ﬂeet vehicle routing with overloads and time windows (HFVROTW), motivated by the fact that real-life scheduling of vehicles involves, to some extent and in addition to time restrictions, some small capacity violations in most distribution scenarios or public transportation in urban or rural areas. Indeed, bus overloads in peak hours, truck overloads enforced by large product demand, or overloads in power and telecommunication networks, are often encountered n Corresponding author. Tel.: þ30 210 8203449; fax: þ 30 210 8828078. E-mail addresses: [email protected] (M.N. Kritikos), [email protected] (G. Ioannou). 0925-5273/$ - see front matter & 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.ijpe.2013.01.020 in real-life but when approaching routing from a research perspective, these facts are always ignored. The HFVROTW can be described as follows: Consider a heterogeneous ﬂeet of vehicles, i.e., comprising vehicles with different capacities, located at a central depot (distribution center or transportation hub). The vehicles are required to serve a set of customers, which are geographically dispersed in the area covered by the depot. Each customer has a known demand and a time window for service. Also, there is a service time associated with each customer, and the distance between each pair of customers is known, as is the distance between all customers and the depot. In the solution of the HFVROTW, vehicles are allowed to carry load over their capacity at a penalty incurred in the total solution cost. Thus, in contrast to the classical vehicle routing problem, the goal of the HFVROTW is to minimize a combined objective of the total distance traveled by vehicles, the ﬁxed costs of vehicles performing service, and the capacity violations of all vehicles included in the ﬁnal schedule. An implicit decision embedded in the problem is the selection of the ﬂeet’s composition, i.e., how many vehicles of each available type (capacity) are selected for service. To our knowledge, research related to the HFVROTW is nonexistent. Most approaches deal either with: (a) the heterogeneous vehicle routing problem (HVRP—i.e., the problem with vehicles of different capacities) or (b) the heterogeneous vehicle routing problem with time windows (HVRPTW—i.e., the previous problem with time windows). Heuristic methods proposed for the HVRP include, as described in Golden et al. (1984), adaptations of the Clarke and Wright savings algorithms, the giant tour partitioning approach, the M.N. Kritikos, G. Ioannou / Int. J. Production Economics 144 (2013) 68–75 matching based savings heuristics, the generalized assignment based heuristic, the sophisticated improvement based heuristic, composite heuristics, and a multi level composite heuristic for the multidepot HVRP. Liu and Shen (1999) were the ﬁrst to tackle the HVRPTW and developed a number of parallel insertions heuristics based on the insertion scheme of Solomon, and embedding in the calculations of the relevant criteria the acquisition costs of Golden et al. (1984). Dullaert et al. (2002) proposed a sequential construction algorithm, extending Solomon’s I1 heuristic with vehicle insertion savings calculations based again on the criteria of Golden et al. (1984). Dondo and Cerda (2007) proposed a 3-phase algorithm for the multi-depot HVRPTW motivated by cluster-based optimization, while Paraskevopoulos et al. (2008) presented a two-phase solution framework relying on a hybridized tabu search integrated within a new reactive variable neighborhood search meta-heuristic algorithm, with very good results. Braysy et al. (2008) presented a deterministic annealing metaheuristic for the HVRPTW, outperforming the results of Liu and Shen (1999), and Braysy et al. (2009) developed a linearly scalable hybrid threshold-accepting and guided local search metaheuristic for solving large scale HVRPTW instances. Finally, Repoussis and Tarantilis (2010) proposed an Adaptive Memory Programming solution approach for the HVRPTW that provides very good results in the majority of the benchmark instances examined. In a relevant research thread, Rochat and Semet (1994) developed a tabu search approach for the HVRPTW, which takes into account the drivers’ breaks and possible accessibility restrictions. Brandao and Mercer (1997) developed also a tabu search for the multi-trip vehicle routing and scheduling problem, in which each vehicle can make several trips per day, while access can be restricted for some vehicles to some customers; the algorithm the authors proposed allows for both weight and volume capacity restrictions on the vehicles. In this work, we address for the ﬁrst time the vehicle routing problem with time windows when the ﬂeet is heterogeneous, i.e., comprises vehicles of different capacities and associated costs, and overloads are allowed up to a pre-speciﬁed bound, at a penalty though embedded in the problem’s objective function. The penalty, which is a measure of the deviation of the actual load from the nominal vehicle capacity, is similar to the one presented by Gheysens et al. (1984), while the capacity bound varies in order to examine a large area of the potential solution space. For the solution of the HFVROTW we propose a simple solution method, i.e., a sequential insertion heuristic, extending the traditional insertion criteria of Solomon (1987), and adapting Golden et al.’s (1984), Dullaert et al.’s (2002), and Liu and Shen’s (1999) ones. The computational results on benchmark problems reinforce our intuition for practical applicability of the proposed approach, with minimal adverse effects on vehicle loads and positive impact on total costs. The remainder of the paper is organized as follows. In Section 2 the HFVROTW is formulated. Section 3 offers the basic contribution of our work, i.e., the way we devise and employ the overload penalty function, the criteria we use within the solution schemes, and the overall solution approach we propose. Section 4 presents an illustrative example based on sample literature instances, and Section 5 includes computational results on benchmark problems. Finally, Section 6 provides our concluding remarks and suggestions for future research. 69 which is a distinct node of the underlying connected graph. Indices i, j and u refer to customers and take values between 2 and n, while index i¼1 refers to the depot; an additional index k counts the vehicles. Vehicles are initially located at the central depot. Each customer i poses a demand qi, requires a service time, si, has a time window [ei, li], and is serviced by exactly one vehicle. There is a cost ckij, (related to the travel time tkij and distance dij) associated with the path from customer i to customer j, using vehicle k. Furthermore, a ﬁxed acquisition cost fk is incurred for each of vehicle k in the routes. Each route originates and terminates at the central depot and must satisfy the time window constraints, i.e., a vehicle cannot start servicing customer i before ei and after li; however, the vehicle can arrive before ei and wait for service. Note that capacity constraints are relaxed in the HFVROTW. Gheysens et al. (1984) presented a mathematical programming formulation for the HVRP. We extend their model to formulate the HFVROTW, the variables of which include: (a) the arrival–departure time to/from customer i, respectively, denoted by ai and pi for each customer I; (b) the vehicle load Qk; (c) he sequence in which vehicles visit customers, xkij, and (d) the activation of a vehicle k, zk. Variables (c) and (d) are deﬁned as follows: 8 > < 1, if vehicle k travels from i to j k ð1Þ xij ¼ and > : 0, otherwise ( zk ¼ 1, if vehicle k is active ð2Þ 0, otherwise Given the above-deﬁned variables the HFVRPOTW can be formulated as follows: T X n X n X Minimize cij xkij þ k ¼ 1i ¼ 1j ¼ 1 T X k¼1 X f k zk þ lk ðQ k C k Þ ð3Þ k A Tn Subject to: n X T X xkij ¼ 1, 8j ¼ 2,3,. . .,n ð4Þ xkij ¼ 1, 8i ¼ 2,3,. . .,n ð5Þ i ¼ 1k ¼ 1 n X T X j ¼ 1k ¼ 1 xkij rzk , 8i,j ¼ 1,2,. . .,n, 8k ¼ 1,2,. . .,T ð6Þ \ n X xk1j r 1 8k ¼ 1,2,. . .,T ð7Þ xki1 r 1 8k ¼ 1,2,. . .,T ð8Þ j¼2 n X i¼2 n X i¼1 xkiu n X xkuj ¼ 0, 8k ¼ 1,. . .,T, 8u ¼ 2,. . .,n ð9Þ j¼1 2. Mathematical model aj Zðpi þ t ij Þð1 xkij ÞM, 8i,j ¼ 1,2,. . .,n, 8k ¼ 1,2,. . .,T ð10Þ The HFVROTW can be stated as follows: Find a set of closed routes, for a ﬂeet of T vehicles with known capacities C1,C2,y,CT, servicing a set of 9L9 1¼ n 1 customers, from a central depot at minimum cost. L is the set of customers including the depot, aj rðpi þ t ij Þð1 xkij ÞM, 8i,j ¼ 1,2,. . .,n, 8k ¼ 1,2,. . .,T ð11Þ ai rpi si , 8i ¼ 1,. . .,n ð12Þ 70 M.N. Kritikos, G. Ioannou / Int. J. Production Economics 144 (2013) 68–75 ei r pi si rli , 8i ¼ 1,. . .,n ð13Þ a1 ¼ 0 n X i¼1 ð14Þ qi ð n X xkij Þ ¼ Q k , 8k ¼ 1,2,. . .,T ð15Þ to Gheysens et al. (1984) for the Heterogeneous VRP, appropriately modiﬁed to address the Heterogeneous VRPTW. The heuristic terminates with a complete HFVROTW solution that satisﬁes all problem constraints. The elements of the proposed solution method are examined in detail below. j¼1 Q k ZC k ð1 þ aÞ, 8k ¼ 1,2,. . .,T xkij A f0,1g, 8i,j ¼ 1,. . .,n, zk A f0,1g, 8k ¼ 1,2,. . .,T and 0 r a r1 8k ¼ 1,2,. . .,T 3.1. Customer selection criteria ð16Þ ð17Þ ð18Þ P In this problem T n ¼ fk= i ¼ 1 qi ð nj¼ 1 xkij Þ 4C k g, which is the set of vehicles that exhibit capacity violations; for these vehicles, we employ parameters lk Z0, that determine a penalty included in the objective function. The values of lk0 s are ﬁxed and they are mathematically deﬁned in Gheysens et al. (1984); for a thorough discussion on parameters lk Z0, the reader is referred to Section 3.2. The objective function (3) models the trade-off between route and vehicle costs, and comprises the penalty for overloads. The ﬁrst term of (3) reﬂects the cost of the routes followed by all vehicles after they depart from and before they return to the depot, as well as the cost of the ﬁrst and last segment of each route. The second term of (3) reﬂects the total vehicle acquisition or set-up cost. The third term is the penalty associated with capacity violations, following the approach of Gheysens et al. (1984). Constraints (4) and (5) ensure that exactly one vehicle enters and departs from every customer and from the depot. Constraint (6) guarantees that no customers are serviced by inactive vehicles, i.e., by vehicles with zk ¼0. Constraints (7) and (8) account for the availability of vehicles by bounding the number of arcs, related to each vehicle k, directly leaving from and returning to the depot, below one, respectively. Constraint set (9) is the typical ﬂow conservation equation that ensures the continuity of each vehicle route. Constraints (10)–(14) are related to time windows and guarantee the feasibility of the schedule for each vehicle. In particular, constraints (10) and (11) ensure that, if customers i and j are consecutive in the schedule of vehicle k, then the arrival time at customer j equals the departure time from customer i, plus the travel time between these two customers; note that M is a large number. In case customers i and j are not serviced by the same vehicle or are not consecutive, constraints (10) and (11) are inactive. Constraints (12) and (13) guarantee that the relationships between arrival time, departure time, and service time with respect to customer i are compatible to the customer’s time window. Constraint (14) sets the departure time from the depot equal to zero, since all routes originate at the depot. Constraint (15) calculates the total load of each vehicle and sets it equal to the relevant variable Qk. Constraint (16) allows the vehicle capacity violation up to a percentage of the total capacity, which is determined by the scheduler through parameter a. Finally, constraints (17) and (18) enforce integrality for the xkij and zk variables, respectively. Pn Let i and j be two consecutive customers in a partially constructed route r, and let u denote an unassigned customer. The function that measures the cost of inserting u between i and j is denoted by c1(i,u,j); it is a composite criterion merging via appropriate weights several sub-metrics and incorporates the overload penalties that will be discussed later in this section. The ﬁrst such metric is the coverage of the time window for the selected customer u (see Ioannou et al., 2001) denoted by c0ij ¼ au eu . Furthermore, c1(i,u,j) includes the time gap c1ij between the latest service time lu of customer u and the vehicle’s arrival time at customer u, caused by the insertion of customer u between customers i and j; this metric expresses the compatibility of the time window of the selected customer with the speciﬁc insertion place in the current route (see also Ioannou et al., 2001 for clariﬁcations) and is given by: c1ij ¼ lu ðau þ su þ diu Þ. In addition, the insertion of u between i and j results in a distance increases c2ij ¼ diu þ duj dij , a metric also included in c1(i,u,j). Metric c3ij ¼ 9C k Q k qu 9 models the deviation from the vehicle’s capacity when customer u is inserted into the route; note that we employ the absolute value for this difference since overloads may cause negative expression of cij 3 . Finally cij 4 , measures the difference of the average ﬁxed cost per unit transferred by vehicle k prior and after the insertion of customer u: c4ij ¼ 9ðf k =Q k Þ½f k =ðQ k þqu Þ9. The vehicle saving insertion metric c5ij is equal to one of the adapted savings concepts deﬁned in Dullaert et al. (2002), i.e., c5ij ¼ ACS, AOOS, AROS. The adapted combined savings (ACS) concept is deﬁned as the difference between the ﬁxed costs of the vehicle capable of transporting the load of the route after and before inserting customer u, f(Qnew) f(Q). To reﬂect the original notion of Golden et al.’s (1984) optimistic opportunity savings, the adapted optimistic opportunity savings (AOOS) concept extends the ACS by subtracting f ðQ new Q new Þ. This is the ﬁxed cost of the smallest vehicle that can service the unused capacity Q new Q new . The adapted realistic opportunity savings (AROS) concept takes the ﬁxed cost of the largest vehicle smaller than or equal to the unused capacity, F 0 ðQ new Q new Þ, into account as opportunity saving. It only does so if a larger vehicle is required to service the current tour after a new customer has been inserted. For a detailed discussion of the aforementioned metrics, refer to Dullaert et al. (2002). Finally, the penalty component c6ij enables the selection of vehicles for which the sum of their ﬁxed cost and capacity penalty is minimized for each route; c6ij is formally deﬁned in the next subsection. Thus, c1(i,u,j) is given by the following relationship: c1 ði,u,jÞ ¼ g1 ðc0ij þc1ij Þ þ g2 c2ij þ g3 c3ij þ g4 c4ij þ g5 c5ij þ g6 c6ij 3. Solution approach To solve the problem modeled in Section 2, we propose a new heuristic method that is based on the insertion framework (Solomon, 1987), and includes two main components in the selection criteria of the non-routed customers: The ﬁrst component is a variant of the criteria proposed by Golden et al. (1984), and also used by Dullaert et al. (2002), Paraskevopoulos et al. (2008), and Liu and Shen (1999). The second component is a penalty function brought into the problem’s objective according ð19Þ In (19), the g weights deﬁne the relative contribution of each individual metric to the overall selection criterion and g1,g2,g3,g4,g5,g6 Z0. To identify the best combination of these weights for the implementation of our algorithm, the following evaluation procedure was used: a. For each g, we have performed an iterative search with a stepincrease of 0.1 (i.e., setting g1 ¼0.1, 0.2, 0.3, y, 0.9, 1.0, then g2 ¼0.1, 0.2, 0.3, y, 0.9, 1.0, and so on) while keeping all the M.N. Kritikos, G. Ioannou / Int. J. Production Economics 144 (2013) 68–75 other g values constant and equal to 0.5 (to simulate some form of equal contribution of all these other g’s). b. For a each of these iterations, we have calculated the value of the objective function and determined the best (minimum) one. c. We examined the sensitivity of the problem solution to the changes in the g-values. Our observations led to the conclusions below: i. The value of the objective function is very sensitive to the changes in the value of parameters g2 and g6; thus we keep this level of step (0.1) in the actual implementation, leading to 10 10 total iterations. ii. The value of the objective function is insensitive to the changes in the value of parameters g3 and g4; thus, we kept just two levels for these parameters (0.5 and 1.0), leading to 2 2 iterations. iii. The value of the objective function is moderately sensitive to the changes in the value of parameters g1 and g5; thus we keep a level of step equal to 0.2 in the actual implementation, leading to 5 5 total iterations. The above iterative scheme was performed in order to approximate some form of local search, as proposed by Dullaert et al. (2002), and increased the computational time of the proposed approach (since a total of 10,000 iterations were embedded within the algorithmic steps) in order to reach better quality solutions. 71 in the application of our algorithm, we further constraint the capacity violation area. This is achieved by employing a factoring metric a that scales the capacity of each vehicle, with a being from 0.1, 0.2 and up to satisfy the yk relationship below. y0k ¼ C k þ a C k , where y0k ryk In this way we are capable of examining multiple scenarios of potential capacity violations, while restricting to the bare minimum the vehicle overloads. 3.3. Criterion for customer insertion into a route In the second phase we can deﬁne the customer selection criterion, i.e., the customer that is best according to the selection criterion c2 ði,u,jÞ ¼ max½c2 ði,u,jÞ, u where c2(i,u,j)¼(m(dou þduo) þsu þF(qu) c1(i,u,j)),m Z0, with su the service time of customer u, m a positive real number, and F(qu) the ﬁxed cost of the smallest vehicle capable of moving a load qu. The above selection criterion ensures that a customer u selected for insertion into a route attempts to maximize the beneﬁt derived from inserting this customer on the partial route being constructed, rather than commencing a new route. 3.2. The penalty component We solve the overloading heterogeneous vehicle routing problem with time windows using a penalty function. The idea is to allow violations in the capacity constraints of some vehicles but to penalize the resulting overloads. Fig. 1 illustrates how the penalty function works. In Fig. 1, parameters yk 1 and yk are the bounds associated with the allowable capacity violations. This means that if a vehicle of nominal capacity of Ck is loaded over this capacity and up to yk, (Ck rQk ryk), then this is allowed at a penalty, determined together with the vehicle’s ﬁxed cost by the relationship: P(Qk) ¼fk þ lk(Qk Ck); a similar deﬁnition holds for yk 1. If the vehicle is loaded up to Ck, while exceeding the bound imposed by the vehicle of smaller capacity (i.e., yk 1 rQk rCk), then the penalty is zero and the following holds: P(Qk)¼fk. Note that the bounds yk are determined by the equation: fk þ lk(yk Ck)¼fk þ 1 (see Gheysens et al., 1984, for further discussion). The values of lk are deﬁned via the a simple parameter y which ensures that yk oCk þ 1 for k¼1, 2, y., T 1, T (the number of vehicle types). y ¼(yk Ck)/(Ck þ 1 Ck), then lk ¼(1/y)(fk þ 1 fk)/ Setting (Ck þ 1 Ck) ¼(fk þ 1 fk)/(yk Ck) for k¼ 1, y, T 1. The penalty component is, thus, deﬁned as c6ij ¼P(Qk) for the active vehicle k for which the current route is constructed. Note that 3.4. Overall solution heuristic To generate a feasible route for the HFVROTW problem, we use the steps of the algorithm presented below: Algorithm OVER-L Step 0: Initialization. Read n, T, Ck, fk, k¼1,2,3y.T, cij, and ei, li, 8i, j ¼2,y, n. Also J¼L. Step 1: Select a ‘‘seed’’ customer to start a route r, ﬁnding the farthest customer from the depot. If there does not exist a nonrouted feasible customer to start a route, go to Step 6. Step 2: Find the feasible non-routed customer u that minimizes the composite criterion c1(i,u,j), i.e. Step 2a: Examine all possible feasible insertions of customer u into the current route. For each feasible insertion, calculate c1(i,u,j). Select insertion location that results in minimum c1(i,u,j) for this customer. Step 2b: Repeat Step 2a for all feasible non-routed customers. Step 2c: Select customer u with maximum c2(i,u,j). Step 3: Insert the selected customer u, to the best insertion location on the current route r (see Steps 2a and 2c). Update the route and set u as a routed customer, decreasing set J. Step 4: If there are non-routed customers that are feasible for insertion into the current route r, return to Step 2; otherwise proceed to Step 5. Step 5: If all customers have been scheduled, go to Step 6. Otherwise, go to Step 1—initiate new route. Step 6: Terminate; output number of routes (active vehicles), sequence of customers visited by each vehicle, total distance (time) and total cost. fk+1 fk Penalty fk-1 Ck-1 yk-1 Ck yk Fig. 1. The shape of the penalty function P(Q). Ck+1 The above algorithm terminates with a complete solution to the HFVROTW. 72 M.N. Kritikos, G. Ioannou / Int. J. Production Economics 144 (2013) 68–75 4. Illustrative example following values of l: To illustrate the mechanics of OVER-L and decipher the ideas behind capacity violations, we provide an example of its application on benchmark data set R101 of Solomon (1987). This data set includes 100 customers, and the best solution reached by heuristics for the heterogeneous vehicle ﬂeet version comprises 24 vehicles (one of type A, eleven of type B, eleven of type C, one of type D, and none of type E, denoted by A1,B11,C11,D1,E0) and has a total cost of 5061 units (see Paraskevopoulos et al., 2008). The vehicle types A, B, C, D and E have capacities 30, 50, 80, 120 and 200 units, and ﬁxed costs 50, 80, 140, 250, 500 units, respectively. Setting y ¼0.5, the penalty function of Section 3 gives the lA ¼ 8050 0:5ð5030Þ lB ¼ 14080 0:5ð8050Þ lC ¼ 250140 0:5ð12080Þ lD ¼ 500250 0:5ð200120Þ Table 1 OVER-L results for R101. Total cargo 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 78 86 112 79 83 75 115 75 59 75 81 82 123 88 39 57 40 50 34 27 Fixed cost plus penalty Fixed cost Fixed cost with violation without violation 140.00 140 140 þ5.5(86 80) ¼173.00 140 250.00 250 140.00 140 140 þ5.5(83 80) ¼156.50 140 140.00 140 250.00 250 140.00 140 80þ 4(59 50) ¼ 116.00 80 140.00 140 140 þ5.5(81 80) ¼145.50 140 140 þ5.5(82 80) ¼151.00 140 250 þ6.25(123 120)¼ 268.75 250 140 þ5.5(88 80) ¼184.00 140 50þ 3(39 30) ¼ 77.00 50 80þ 4(57 50) ¼ 108.00 80 80.00 80 80.00 80 50þ 3(34 30) ¼ 62.00 50 50.00 50 Total cost 2620 ¼ 30 0:520 ¼ ¼ 60 0:530 ¼4 ¼ 110 0:5ð40Þ ¼ 250 0:580 3 1 ¼3 ¼ 5:5 ¼ 25 4 ¼ 6:25 lE ¼ lD Applying OVER-L to the data, we reach the solution shown in Table 1, with 20 active vehicles of types A3, B4, C10,D3,E0. Table 1 shows the vehicle ﬁxed cost in the solution where capacity violations are allowed, and also provides the respective ﬁxed costs if violations are eliminated by activating larger capacity vehicles. It also provides the total load carried by each vehicle serving the route, as well as the ﬁxed cost incorporating the penalty component. Note that the results of Table 1 have been obtained without using the a parameter of the customer insertion rule; we rather rely directly on the data of Gheysens et al., (1984) Table 3 Results for C101. 140 250 250 140 250 140 250 140 140 140 250 250 500 250 80 140 140 140 80 50 3720 Customers 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Distance Waiting time 82,79,77,72,71 120.96 91,88,87,84,83,74,78,80,81 127.71 56,64,54,57,59,61,60,51,53,50,48 163.26 99,97,96,95,93,85 103.62 6,4,8,9,11,94 103.09 21,34,36,38,39 89.06 19,20,16,17 90.67 33,32,30,40,37,35 115.65 14,18,63,75,12,98,101,100,2,76 201.62 58,55,41,45,46 88.94 44,25,26,28,31,15,13 112.94 68,66,73,86,89,90,92 89.72 43,42,47,62,49,52,23 105.78 29,65,69,67,70 79.27 10,7,5,3,22 61.81 27,24 31.81 Total 1685.89 0 0 0 0 0 71.17 15 224.73 9.20 114.43 47.20 245.61 193 0 12.92 17 950.26 Vehicle cost Route cost 300 300 800 300 300 300 300 300 800 300 300 300 300 300 300 300 5800 420.96 427.71 963.26 403.62 403.09 460.23 405.67 640.38 1010.82 503.37 460.14 635.33 598.78 379.27 374.73 348.81 8436.15 Table 2 Additional OVER-L results for R101. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Customers Overload Distance Waiting time Vehicle cost Route cost 66,72,10,35,36,71 64,65,50,33,78 40,24,68,56,5,26 15,45,39,44,92,101 37,48,20,9,47,18,94 63,12,91,67,2,81 93,43,99,17,87,38,61,90,59 34,30,52,21,49 28,70,79,4,55,25 60,96,16,42,57,75 46,84,62,85,97,14 73,76,74,23,98 32,31,82,51,69 29,13,77,80,27 53,89,8,11 6,83,19,7 3,88,58 100,86 22,41,54 95 0 (78) 6 (86) 0 (112) 0 (79) 3 (83) 0 (75) 0 (115) 0 (75) 9 (59) 0 (75) 1 (81) 2 (82) 3 (123) 8 (88) 9 (39) 7 (57) 0 (40) 0 (50) 4 (34) 0 (27) Total 138.46 142.42 125.10 111.69 136.64 129.31 116.27 131.96 102.14 97.37 88.57 82.77 89.71 65.43 66.33 68.58 55.48 46.17 36.28 24.08 1854.76 13.57 25.92 22.96 47.35 11.61 0.79 17.30 14.14 7.99 26.03 28.76 18.60 5.17 5.07 12.47 7.40 25.93 0.00 5.93 0.00 296.99 140 140 250 140 140 140 250 140 80 140 140 140 250 140 50 80 80 80 50 50 2620 292.03 308.34 398.06 299.04 288.25 270.10 383.57 286.10 190.13 263.40 257.33 241.37 344.88 210.50 128.80 155.98 161.41 126.17 92.21 74.08 4771.75 M.N. Kritikos, G. Ioannou / Int. J. Production Economics 144 (2013) 68–75 Table 2 provides additional data concerning the output of OVER-L, i.e., the subset of customers and the level of overload of each route (total load in parenthesis), the total route distance, the overall waiting time of each route, the vehicle ﬁxed cost and the total route cost; the latter includes the vehicle ﬁxed cost, the route distance that is equivalent to route time, and the waiting time at each customer except the ﬁrst. From Table 2, it is evident that small violations of the vehicle capacities result in signiﬁcantly lower ﬁxed vehicle costs, a fact that supports our claim that the HFVROTW is a worth studying problem. The total cost of the solution determined by OVER-L is 4771.75, as shown in Table 2. This value is lower than the best solution found in the literature for the case where violations are not allowed, even by complex and intelligent meta-heuristics (see Paraskevopoulos et al., 2008). Thus, the controlled violations of the capacity of vehicles results in reduced total costs for routing and scheduling problems with minimal capacity violations. The results of Table 3, referring to data set C101 of Solomon (1987), further reinforce and conﬁrm that even for very restrictive Table 4 OVER-L vis-a -vis Liu and Shen’s heuristic. Instance Liu and Shen OVER-L (%) DEV R101 R102 R103 R104 C101 C102 C103 RC101 RC102 RC103 5061.00 5013.00 4772.00 4455.00 9272.00 8433.00 8033.00 5687.00 5649.00 5419.00 4687.00 4655.00 4484.00 4165.00 7286.00 7553.00 7722.00 5176.00 5175.00 5009.00 7.38 7.14 6.03 6.50 21.41 10.43 3.87 8.98 8.39 7.56 73 problems such as C101, where time windows in combination with service time play the most critical role in determining feasible solutions, the proposed approach is superior to other methods. Speciﬁcally, for data set C101, OVER-L derives a solution with 16 vehicles and a total cost of 8436.15; this solution is almost 10% better (lower) than the one derived by Paraskevopoulos et al. (2008), which has a cost of 9272. This is directly attributed to the small capacity violations that can reduce signiﬁcantly the ﬁxed vehicle costs. Note that the results of Table 3 are obtained using a a parameter equal to 0.1 (a ¼0.1), while one could expect much larger cost reductions if the HFVROTW was solved via complex meta-heuristics. 5. Computational results We have executed numerical experiment of the proposed algorithm on the benchmark data sets of Solomon (1987)—100 customer instances R, C, and RC proposed for the original VRP with time windows. All experiments were performed on a PC equipped with a Pentium Dual Core T4200 running at 2.0 GHz. We compare our heuristic’s performance to Liu and Shen (1999). As we have already mentioned the objective is to achieve cost reductions with minimal and controlled violations in the capacity of vehicles. Liu and Shen (1999) use the distance travelled, the waiting time (excluding the waiting time for the ﬁrst customer) and the vehicle costs to measure the solution quality. Our approach reduces the total delivery cost as deﬁned by Liu and Shen (1999), allowing a small percentage of capacity overloads. Table 4 illustrates the performance of OVER-L vis-a -vis the Liu and Shen (1999) approach for the HFVROTW problem, in terms of the total cost. The results show that this metric is substantially improved by OVER-L. Note that for all the data sets reported in Table 5 Results for R1 data set. Instance Vehicles R101 R102 R103 R104 R105 R106 R107 R108 R109 R110 R111 R112 A3B4C13 A4B3C5D5E1 A1B1C4D8 B1C2D9 B2C8D5 A2B1C3D8 A2C2D9 A1C1D10 DC4D8 A1C4D8 B1C1D10 B 1 C 2 D9 Distance Cost w/o violation Cost with violation FC w/o violation FC with violation % improvement on FC cost % capacity violation CPU times (sec) 1920.84 1793.33 1613.73 1372.01 1659.43 1561.11 1426.07 1343.42 1471.74 1498.86 1483.42 1336.47 4687.30 4655.89 4484.37 4165.56 4247.94 4200.48 4151.17 4082.19 4141.99 4130.64 4107.69 3963.47 3470 4529 4410 5140 200 4990 5020 5050 4970 4800 5080 5140 2490 2640 2690 2610 2530 2600 2630 2690 2640 2610 2610 2610 28.24 41.70 39.00 49.22 39.76 47.89 47.60 46.73 46.88 45.62 48.62 49.22 5.8 7.8 7.9 11.5 10.4 10.6 11.3 10.6 9.1 10.4 12.2 11.5 67.39 108.52 160.14 253.63 93.02 140.65 193.47 252.48 136.07 183.24 173.60 277.17 Cost with violation FC w/o violation FC with violation % improvement on FC cost % capacity violation CPU times (sec) 7286.64 7553.65 7722.71 7315.52 7093.56 7413.55 7067.80 7043.14 7032.74 11300 12050 10750 11550 11050 11300 11300 11300 11050 5600 5000 5600 5800 5400 5200 5600 5600 5800 50.44 58.50 47.90 50.21 51.13 53.98 50.44 50.44 47.51 13.8 15.4 12.1 11.6 14.9 14.9 16.0 16.0 14.9 114.65 158.52 219.15 278.56 121.05 136.43 133.70 169.65 234.83 5667.30 6545.89 6204.38 6695.56 5917.94 6590.48 6541.17 6442.20 6741.99 6320.64 6577.69 6493.47 (86) (114) (116) (168) (153) (156) (166) (155) (133) (153) (176) (168) Table 6 Results for C1 data set. Instance Vehicles Distance Cost w/o violation C101 C102 C103 C104 C105 C106 C107 C108 C109 A8B4 A14B1 A8B4 A6B5 A10B3 A12B2 A8B4 A8B4 A6B5 1338.81 1902.34 1461.12 1313.58 1349.31 1557.71 1411.74 1348.95 1232.74 12986.65 14603.65 12872.72 13067.57 12743.56 13513.56 12767.81 12743.15 12282.74 (250) (280) (220) (210) (270) (270) (290) (290) (270) 74 M.N. Kritikos, G. Ioannou / Int. J. Production Economics 144 (2013) 68–75 Table 7 Results for RC1 data set. Instance Vehicles RC101 RC102 RC103 RC104 RC105 RC106 RC107 RC108 A3B10C3D1 A3B5C6D1 A4B3C7D1 A1B8C2D2 A3B9C4D1 A1B8C6 A1B7C5D1 A1B6C5D1 Distance Cost w/o violation Cost with violation FC w/o violation FC with violation % improvement on FC cost % capacity violation CPU times (sec) 2047.47 1890.71 1696.47 1786.13 1904.42 1808.04 1821.07 1678.23 5176.63 5175.98 5009.03 4590.82 5187.65 4941.00 4881.43 4596.15 5010 4470 4680 3900 5070 4949 4410 3960 3030 3180 3240 2760 3180 3060 3060 2910 65.34 40.56 44.44 41.30 59.43 61.76 44.11 36.08 9.8 7.6 6.3 4.8 9.5 10.2 5.6 4.0 77.42 109.87 144.52 193.91 133.33 108.03 141.35 179.19 7156.64 6465.98 6449.03 5730.82 7077.65 6831.00 6231.43 5646.16 Table 4, the maximum violation is less than 10% of the nominal capacity in the majority of the cases, while the parameters of our algorithm were set to y ¼0.5 (to follow Gheysens et al., 1984) and a ¼0.2. The results for all R1, C1 and RC1 data set are summarized in Tables 5–7. Speciﬁcally, we present the route distance, the cost without violations and the cost with violations, the ﬁxed cost (FC) without violations, the ﬁxed cost (FC) with violation, the percentage improvement in the ﬁxed and total cost due to the capacity violations, and the CPU times. From the results of Tables 5–7, we can securely infer that there is indeed cost reduction in all routing instances with capacity violations that are quite insigniﬁcant. This is exactly the goal of the proposed approach and the fact that it can be demonstrated in most Solomon’s data sets, which were not developed for testing such problems, is very encouraging. Furthermore, the CPU times are much shorter from those of the metaheuristic of Braysy et al. (2008). The latter is demonstrated in the aggregate data concerning CPU times that Table 8 provides. Note that we report only aggregate values since Liu and Shen (1999) report only such CPUtime results; in addition, for fair comparison, we include in parentheses the CPU’s used to run the experiments for each method. Thus, when comparing our CPU times to Liu and Shen (1999), one must consider the fact that the computer used by Liu and Shen (1999) is slower by more than one order of magnitude to the one used for our tests. Finally, Table 9 extends the results of Table 4 for all Solomon’s data sets included in our computational experiments. The values for a–c are adopted from Liu and Shen (1999) and are given in Table 10; these values refer to cost and capacity equivalences. From the results of Table 9, one can see that our approach improves the total cost from 2% (for data set RC1c) to almost 10% (for data set C1b) by small capacity violations. This is very important since it has been demonstrated for a very large set of data (36 data sets for R1, 27 data sets for C1, and 24 data sets for RC1). Finally, to provide a vis-a -vis list of results from metaheuristics and OVER-L, we include in Table 11 the total cost of OVER-L and the total cost of the best metaheuristic solution as reported by Repoussis and Tarantilis (2010) for the data set R1 of Solomon (1987). The results show that this metric is improved by OVER-L, as was the case for the Liu and Shen heuristic. Note that for all the data sets reported in Table 11, the maximum violation is less than 15% of the nominal capacity (a fact that justiﬁes the differences with Table 4). (170) (132) (110) (84) (164) (177) (98) (70) Table 8 CPU time comparisons (aggregate)—times in sec. Data set Liu and Shen (Pentium II, 233 MHz) Braysy et al. (AMD Athlon 2600 þ , 2.13 GHz) OVER-L (Pentium Dual Core T4200, 2.0 GHz) R1 (36) C1 (27) RC1 (24) 175.83 144.00 155.67 787.99 641.28 536.66 152.85 148.80 128.76 Table 9 OVER-L vis-a -vis Liu and Sen’s heuristic. Data set Liu and Sen OVER-L (%) DEV R1a(12) R1b(12) R1c(12) R1(36) C1a(9) C1b(9) C1c(9) C1(27) RC1a(8) RC1b(8) RC1c(8) RC1(24) 4562.00 2149.00 1788.00 2833.00 8042.00 2626.00 1870.00 4179.33 5184.00 2235.00 1849.00 3089.33 4251.00 2080.00 1741.00 2690.00 7281.00 2544.00 1828.00 3884.33 4944.00 2187.00 1813.00 2981.33 6.81 3.32 2.62 5.04 9.50 3.12 2.21 7.05 4.62 2.15 1.95 3.49 Table 10 Liu and Shen (1999) heterogeneous vehicle ﬂeet characteristics for the HFVROTW (subclasses a, b, c for R1, C1, RC1 data set). Vehicles A B C D E Capacity 30 50 80 120 200 R1a 50 80 140 250 500 R1b 10 16 28 50 100 R1c 5 8 14 25 50 Vehicles A B C Capacity 100 200 3009 C1a 300 800 1350 C1b 60 160 270 C1c 30 80 135 Vehicles A B C D Capacity 40 80 150 200 RC1a 60 150 300 450 RC1b 12 30 60 90 RC1c 6 15 30 45 6. Conclusions In this paper, we addressed for the ﬁrst time the vehicle routing problem with time windows when the ﬂeet is heterogeneous, i.e., comprises vehicles of different capacities and associated costs, and overloads are allowed up to a pre-speciﬁed bound, at a penalty embedded in the problem’s objective function. For the solution of the HFVROTW we developed a sequential insertion heuristic, extending the traditional insertion criteria of Solomon (1987), and adapting M.N. Kritikos, G. Ioannou / Int. J. Production Economics 144 (2013) 68–75 75 References Table 11 OVER-L vis-a -vis best metaheuristic for HFVRTW. Data set Best metaheuristic solution OVER-L R101 R102 R103 R104 R105 R106 R107 R108 R109 R110 R111 R112 4536.40 4348.92 4119.04 3981.28 4229.67 4118.43 4031.16 3962.20 4052.21 3995.18 4016.19 3947.30 4242.51 4294.81 4018.14 3831.11 3815.21 3793.39 3831.93 3695.75 3793.19 3762.34 3722.37 3689.69 Golden et al.’s (1984), Dullaert et al.’s (2002), and Liu and Shen’s (1999) ones. The computational results on benchmark problems showed that competitive results are obtained for problems examined and this reinforced our intuition for practical applicability of the proposed approach, with minimal adverse effects on vehicle loads and positive impact on total costs. Future research directions include the development of metaheuristics for the relaxed HVRPTW, hybridization of clever heuristics with complex search methods, examination of non-linear penalty functions, and the study of the effect of very small capacity violations (white noise) to the overall solution quality of the problem. Acknowledgements The authors are indebted to the two anonymous reviewers for their useful comments and suggestions that have helped us improve the original manuscript. Brandao, J., Mercer, A., 1997. A tabu-search algorithm for the multi-trip vehicle routing and scheduling problem. European Journal of Operational Research 100, 180–191. Braysy, O., Dullaert, W., Hasle, G., Mester, D., Gendreau, M., 2008. An effective multi-restart deterministic annealing metaheuristic for the ﬂeet size and mix vehicle routing problem with time windows. Transportation Science 42 (3), 371–386. Braysy, O., Porkka, P.P., Dullaert, W., Repoussis, P.P., Tarantilis, C.D., 2009. A well scalable metaheuristic for the ﬂeet size and mix vehicle routing problem with time windows. Expert System and Applications 36 (4), 8460–8475. Dondo, R., Cerda, J., 2007. A cluster-based optimization approach for the multidepot heterogeneous ﬂeet vehicle routing problem with time windows. European Journal of Operational Research 176, 1478–1507. Dullaert, W., Janssens, G.K., Sorensen, K., Vernimmen, B., 2002. New heuristics for the ﬂeet size and mix vehicle routing problem with time windows. Journal of the Operational Research Society 53, 1232–1238. Gendreau, M., Tarantilis, C.D., 2010. Solving large-Scale Vehicle Routing Problems with Time Windows: The State-of-the-Art. CIRRELT-2010-04. Université de Montreal, Canada. Gheysens, F., Golden, B., Assad, A., 1984. A comparison of techniques for solving the ﬂeet size and mix vehicle routing problem. OR Spektrum 6, 207–216. Golden, B., Assad, A., Levy, L., Gheysens, F., 1984. The ﬂeet size and mix vehicle routing problem. Computers and Operation Research 11 (1), 49–66. Golden, B., Raghavan, S., Wasil, E., 2008. The Vehicle Routing Problem: Latest Advances and New Challenges. Springer, NY. Ioannou, G., Kritikos, M., Prastacos, G., 2001. A greedy look-ahead heuristic for the vehicle routing problem with time windows. Journal of the Operational Research Society 52, 523–537. Kritikos, M., Ioannou, G., 2010. The balanced cargo vehicle routing problem with time windows. The International Journal of Production Economics 123 (1), 42–51. Liu, F.-H., Shen, S.-Y., 1999. The ﬂeet size and mix vehicle routing problem with time windows. Journal of the Operational Research Society 50, 721–732. Paraskevopoulos, D.C., Repoussis, P.P., Tarantilis, C.D., Ioannou, G., Prastacos, G.P., 2008. A reactive variable neighbourhood tabu search for the heterogeneous ﬂeet vehicle routing problem with time windows. Journal of Heuristics 14 (5) 425–255. Repoussis, P.P., Tarantilis, C.D., 2010. Solving the ﬂeet size and mix vehicle routing problem with time windows via adaptive memory programming. Transportation Research Part C 18 (5), 695–712. Rochat, Y., Semet, F., 1994. A tabu-search approach for delivering pet food and ﬂour in Switzerland. Journal of the Operational Research Society 45, 1233–1246. Solomon, M.M., 1987. Algorithms for the vehicle routing and scheduling problems with time windows constrains. Operations Research 35, 254–265.